XInfer.AI builds independent LLM judge to catch AI fluency masking falsehood

XInfer.AI builds AI assistants that explain diamonds to shoppers in a two-voice spoken conversation: a host asks questions, a gemologist answers. The stone's lab certificate from GIA or IGI—the independent laboratories that grade the world's diamonds—provides the evidence. Thousands of dollars and a high-consideration purchase make this an unforgiving test bed for generative AI. Nearly every claim can be checked. This means overselling is not just tacky; it is provably false.

The problem emerged during development. The assistant produced: 'At this clarity grade, you won't see a thing with the naked eye.' The grade, SI1, is real, printed on the lab report. The promise is invented. A clarity grade is assigned under 10× magnification and does not resolve what a naked eye will see. Two SI1 stones can look entirely different. The sentence is fluent, helpful-sounding, and false—a sentence that should keep a team up at night because no simple rule would catch it. It contains no banned word, no hallucinated stone, no wrong number. It breaks no rule written in advance. Fluency and truth are different variables, and a language model optimizes the first.

To fix this, XInfer.AI built a three-layer enforcement system. The first layer prevents: code inside the pipeline rejects any follow-up question that names no available evidence path, so the model never gets a chance to improvise an answer. The second layer catches: a deterministic validator runs regex tests over the finished script—a hard-fail lexicon of words like 'amazing,' 'must-have,' 'investment,' and 'trust me.' If a banned phrase appears, the script fails. Cheap, fast, absolute. The third layer judges: what remains is the dangerous class—plausible-but-unsupported claims like the naked-eye sentence. Checking requires reading both the sentence and the lab report. At scale, the reader has to be a language model.

The judge was built differently from typical LLM ratings. Instead of a score from one to ten—an opinion with a number attached, unauditable—it returns a triage. For every claim beyond restating the input data, it answers typed questions: Is this about this stone or diamonds in general? If about this stone, does the evidence establish it? If general, is it accurate gemology? Lab-grown and natural diamonds are chemically identical—category teaching, permitted and wanted. 'This stone will look white to the eye'—an item claim, must trace to evidence. A real violation is only the intersection: about this stone, and unsupported. The judge also flags misattribution. If the assistant says 'we grade this VS1,' every word is true but the sentence violates the rule—it credits the lab's work to the house. The judge reports the claims it cleared too, with classifications visible, and is forbidden from inflating the violation count. This is the whole point: a verdict showing every ruling with quotes and evidence paths is an audit trail you can re-check. When the judge is wrong, you see exactly where.

The judge itself is an LLM, an unreliable narrator auditing another. Independence had to be constructed. The writers ran on Claude Sonnet 5. The judge's seat was first held by GPT-5.5, then Claude Opus 4.8—a different model from the writers—and the discipline held across the change. The judge never grades its own output; that would embed a conflict of interest. It also never touches production; it costs about thirty cents per diamond (four re-reads of all evidence), which is trivially cheap in development but pointlessly expensive at runtime where deterministic gates already stand. A judge is a development instrument to find what to fix, not a content filter.

The team calibrated the judge by treating it as an instrument. Verdicts were schema-checked to catch malformed outputs; the harness retried bad verdicts. The same stone judged twice did not score the same—one diamond scored fifteen violations on one run and six to eight on reruns. The discipline was boring: average at least two runs before believing any comparison. False positives were treated as calibration data. The team had banned the word 'flaw' but the assistant said 'not a flaw; just part of why the rating sits at four'—the most honest sentence, failed by regex. The ban was narrowed to permit 'not a flaw' as correction while blocking 'no flaws in this stone' as perfection claim. The exemption got its own tests.

Three decisions that AI teams normally settle by taste, seniority, or vendor reputation got settled by number. Which writer model to use: Claude Haiku 4.5 averaged about ten real violations per diamond; Claude Sonnet 5 averaged 0.8—a twelvefold honesty gap on identical inputs and prompts. Sonnet also scored higher on specificity, so honesty did not cost richness. Every arithmetic blunder in the corpus came from the small model, including one run where it declared one stone '59 carats more' than another when the real gap was 0.13 carats—and it had turned a price difference into a weight difference. The comparison was repeatable, auditable, and measured.

The core insight of XInfer.AI's approach is that fluency and truth are orthogonal. A language model optimizes for sounding helpful and coherent, not for factual accuracy—especially when the claim is plausible and contains true sub-components (like a real diamond clarity grade, SI1, paired with a false consequence). No set of advance rules can catch every such sentence, because the falsehood lives in the relationship between the text and the evidence, not in the text alone. This is why content filters fail: they catch obvious violations (banned words, hallucinated numbers) but miss the dangerous class—claims that are internally coherent and sound authoritative while being unsupported. The company's response was to separate the writing task from the judging task and to assign the judge a different, independent model. The judge's workflow also embeds a key discipline: verdicts must be auditable. Instead of returning a score (which demands trust), the judge lists every claim it evaluated, names the category and evidence path, and shows the false positives it found—turning a black-box opinion into a transparent audit trail. This design choice extends to the boundaries between the three gates. Words like 'flawless' and 'guaranteed' were deliberately not blocked by regex, even though they appear in a banned-word list, because in diamond grading they have literal meanings. The regex gate would have caught the most honest uses ('It is not flawless' as a factual statement about an SI1 stone). The judge, by contrast, can read context and distinguish a marketing claim from an accurate statement. Finally, the discovery that Claude Sonnet 5 outperformed Claude Haiku 4.5 by a factor of twelve on honesty—without sacrificing specificity—is an example of how measurement settles arguments that normally devolve into vendor preference or seniority. The same stones, the same prompts, the same judge: the comparison is repeatable and auditable.

No comments yet. Be the first to share your thoughts!